CN1188835C - System and method for reducing noise - Google Patents

Abstract

A noise suppression system and method for a speech processing system (108). A gain estimator (220) determines the gain and noise suppression level for each input signal frame and then sets the gain to a predetermined minimum value. If there is speech in the frame, the adjuster (224) determines a gain factor for each channel of the predefined set of frequency channels. For each channel, the gain factor is a function of the speech SNR within the channel. The channel SNR is generated by the SNR estimator (210b) based on the channel energy estimate provided by the energy estimator (206b) and the channel noise energy estimate provided by the noise energy estimator (214b). The noise energy estimator (214b) updates the estimate of the speech-free frame periods determined by the speech detector (208).

Description

Noise suppression system and method

field of invention

The present invention relates to speech processing. In particular, the present invention relates to noise suppression systems and methods for speech processing.

Background technique

Digital transmission of voice is becoming common, especially in cellular telephone and personal communication systems (PCS) applications. This has generated interest in improving speech processing techniques. One area that is improving is noise suppression technology.

Noise suppression in speech communication systems generally improves the overall quality of the desired audio signal by filtering out ambient background noise from the desired speech. Such speech enhancement techniques are especially needed in environments with unusually high ambient background noise, such as airplanes, moving vehicles, or noisy factories.

One noise suppression technique is spectral subtraction or spectral gain correction. With this method, an input audio signal is divided into frequency channels, and specific frequency channels are attenuated according to their noise energy. The background noise estimate for each frequency channel is used to generate a signal-to-noise ratio (SNR) for speech within the channel, and the SNR is used to calculate a gain factor for each channel. The gain factor then determines the specific channel attenuation. The attenuated channels are recombined to produce a noise-suppressed output signal.

In certain applications involving high background noise environments, most noise suppression techniques are significantly limited in performance. An example of such an application is a car speakerphone option for a cellular mobile communication system. This speakerphone option provides hands-free operation for the driver of the vehicle. Hands-free headsets are typically located far away from the user (eg mounted on a helmet). Headsets farther away have poorer SNR for signals transmitted to land-based parties due to road and wind-induced noise. While speech received at the ground-based end is usually clear, the constant exposure to this background noise often increases listener fatigue.

For a noise suppression system to work properly, it is important to accurately determine the SNR of speech. However, due to limitations of currently used noise detectors, it is difficult to accurately determine the SNR of speech signals. Spectral subtraction techniques update background noise estimates when speech is not present. When speech is not present, the measured spectral energy is attributed to noise, and the noise estimate is updated based on the measured spectral energy. Therefore, in order to obtain an accurate estimate of the noise energy to calculate the SNR, it is important to distinguish periods of speech presence from periods of speech absence.

An exemplary speech detection technique uses a speech metric calculator to perform noise update decisions. The speech metric is a measure of the overall speech-like properties of the channel energy. First, the raw SNR estimates are used to index the speech metrics table to obtain speech metrics for each channel. The individual channel speech metrics are summed to produce an energy parameter, which is compared to a background noise update threshold. If the sum of the speech metrics is equal to or greater than a threshold, the signal is said to contain speech. If the sum of the speech metrics is less than a threshold, the input frame is considered noisy and the background noise update is done. But in the case of high background noise, sudden background noise, or gradually increasing noise sources, the SNR measurement will be large, resulting in a higher speech metric, preventing the update of the noise estimate.

A further improvement to the Speech Metric Calculator technique is the measurement of channel energy deviation. The method assumes that noise has constant spectral energy over time, while speech has varying spectral energy over time. The channel energy is thus integrated over time, and speech is detected if there is a large channel energy deviation, and noise is detected if there is only a small channel energy deviation. A speech detector that measures deviations in channel energy will detect sudden increases in noise. But the channel energy bias method provides inaccurate results when the input speech signal energy is constant. And for the case where the noise source is gradually increasing, the change of the input energy will lead to a large energy deviation, which will prevent the noise estimate from being updated even if it needs to be updated.

In addition to an accurate speech detector, a speech suppression system must properly adjust the channel gain. Channel gain should be adjusted to suppress noise without sacrificing speech quality. One method of channel gain adjustment is to calculate the gain as a function of the total noise estimate and the SNR of the speech signal. In general, an increase in the total noise estimate results in a decrease in the gain factor for a given SNR. A reduced gain factor indicates a larger attenuation factor. This technique applies a minimum gain value to prevent excessive attenuation of the channel gain when the total noise estimate is very large. By utilizing a hard-clamped minimum gain value, a compromise is found between noise suppression and speech quality. When the inlay is lower, noise suppression is improved but speech quality is degraded. When the clamping is higher, the noise suppression is worse but the speech quality is improved. US Patent No. 4811404 discloses a method and apparatus for suppressing background noise in a high background noise environment. This US patent includes an increasing signal-to-noise ratio (SNR) threshold mechanism to reduce background noise due to offsetting the gain increase of the gain table until a certain SNR threshold is reached.

In order to provide an improved noise suppression system, limitations of current techniques for speech detection and channel gain calculation need to be addressed. These problems and disadvantages are solved by the present invention in the following manner.

Contents of the invention

The present invention is a noise suppression system and method for a speech processing system. It is an object of the present invention to provide a speech detector for determining the presence or absence of speech in an input signal. To accurately determine the signal-to-noise ratio (SNR) of speech, a reliable speech detector is required. When it is judged that speech does not exist, the input signal is considered to be completely a noise signal, and the noise energy can be measured. The noise energy is then used to determine the SNR. Another object of the present invention is to provide an improved gain determination unit to suppress noise.

In accordance with the present invention, a noise suppression system includes a speech detector for determining the presence or absence of speech within a frame of an input signal. Speech can be judged from the SNR measure of the speech in the input signal. The SNR estimator estimates the SNR based on the signal energy estimate produced by the energy estimator and the noise energy estimate produced by the noise energy estimator. The voice can also be judged according to the coding rate of the input signal. In a variable rate communication system, each incoming frame is assigned an encoding rate selected from a set of preset rates based on the content of the incoming frame. Typically, the rate depends on the level of speech activity, so frames containing speech will be assigned a higher rate, and frames that do not contain speech will be assigned a lower rate. Furthermore, the speech may be judged based on one or more pattern measurements characterizing the input signal. If it is determined that there is no speech in the input frame, the noise energy estimator updates the noise energy estimate.

A channel gain estimator determines the gain for a frame of the input signal. If there is no speech in the frame, the gain is set to the preset minimum value. Otherwise, the gain is determined based on the frequency content of the frame. In a preferred embodiment, a gain factor is determined for each set of predefined frequency channels. For each channel, the gain is determined based on the in-channel speech SNR. For each channel, the gain is defined using a function appropriate to the characteristics of the frequency band in which the channel resides. In general, for a predefined frequency band, the gain is set to increase linearly with increasing SNR. In addition, the minimum gain of each frequency band can be adjusted according to the environmental characteristics. For example a user selectable minimum gain may be implemented. The SNR of the channel is determined based on the channel energy estimate generated by the energy estimator and the channel noise energy estimate generated by the noise energy estimator. The gains of the signals in the different channels are adjusted using the gain factors, and the gain-adjusted channels are combined to produce a noise-suppressed output signal.

According to the present invention, there is provided a noise suppressor for suppressing background noise causing a signal, comprising: a signal-to-noise ratio estimator for generating a channel SNR estimate for a first predefined frequency channel group of said audio signal; a gain estimate An estimator for generating a gain factor for each of the frequency channels according to a corresponding one of the channel SNR estimators, wherein the gain factor is obtained by using a gain function defined as an SNR increasing function; gain adjustment a device for adjusting the gain level of each of said frequency channels based on one of said corresponding gain factors; and a speech detector for determining the presence of speech in said audio signal, wherein said speech detector utilizes another signal A noise ratio estimator and rate decision unit determines a set of variable rate coding rates for said audio signal to detect the presence of speech.

According to the present invention, there is also provided a method for suppressing background noise of an audio signal, comprising the steps of: transforming said speech signal into a frequency representation of said audio signal; detecting a coding rate associated with said audio signal; Determine whether speech is present in the audio signal at the encoding rate of the audio signal; generate channel signal-to-noise ratio (SNR) estimates for a predefined set of frequency channels represented by the frequency; determine if speech is present in the audio signal for each Gain factors for each of the frequency channels, where a gain function is defined for each of a set of frequency bands, and for each of the frequency bands a gain factor that increases with increasing SNR is defined, so for each of the frequency channels , a channel gain factor determined from a gain function ranging over a frequency band containing frequency channels; adjusting a gain level for each of said frequency channels according to said corresponding channel gain factor; and inverse transforming said gain-adjusted frequency representation to produce noise-suppressed audio Signal.

Brief description of the drawings

The features, objectives and advantages of the present invention can be further understood by the description of the present invention in the following drawings, wherein the same parts are represented by the same reference numerals in the accompanying drawings, wherein:

Figure 1 is a block diagram of a communication system utilizing a noise suppressor;

Fig. 2 is a block diagram according to the noise suppressor of the present invention;

FIG. 3 is a graph of frequency-based gain factors for achieving noise suppression in accordance with the present invention; and

FIG. 4 is a flowchart of an embodiment of processing steps in noise suppression performed by the processing unit of FIG. 2 .

Best Mode of Carrying Out the Invention

In voice communication systems, noise suppressors are often used to suppress unwanted ambient background noise. Most noise suppressors perform suppression by estimating the background noise characteristics of the input data signal in one or more frequency bands and subtracting the estimated average from the input signal. The estimate of average background noise is updated during periods of no speech. Noise suppressors require accurate judgment of background noise levels for proper operation. In addition, the level of noise suppression must be properly adjusted according to the speech and noise characteristics of the input signal. These requirements are addressed by the noise suppression system of the present invention.

Figure 1 shows an exemplary speech processing system 100 according to the present invention. System 100 includes headset 102 , A/D converter 104 , speech processor 106 , transmitter 110 and antenna 112 . Headset 102 may be located within the cellular telephone along with the other elements of FIG. 1 . The headset 102 may also be a hands-free headset of the speakerphone option of a cellular communication system. A car speakerphone kit is sometimes called a carkit. Noise suppression is particularly important where the headset 102 is part of a car kit. Since the hands-free headset is generally located at a certain distance from the user, the speech SNR of the received sound signal is always poor due to road and wind.

Referring to FIG. 1 , a headset 102 receives an input audio signal containing speech and/or background noise. The input audio signal is converted by the earphone 102 into an electro-acoustic signal represented by the term s(t). The electro-acoustic signal may be converted from an analog signal to pulse code modulation (PCM) samples by an analog-to-digital converter 104 . In the exemplary embodiment, the PCM samples are output by A/D converter 104 at 64 kbps and are represented by signal s(n) as shown in FIG. 1 . The digital signal s(n) is received by a speech processor 106 which includes, among other things, a noise suppressor 108 . Noise suppressor 108 suppresses noise in signal s(n) according to the invention. In automotive applications, the noise suppressor 108 determines the level of background ambient noise and adjusts the signal gain to attenuate the effects of such ambient noise. In addition to noise suppressor 108, speech processor 106 typically includes a speech encoder or vocoder (not shown), which compresses speech by extracting parameters related to the production model of the human voice. Speech processor 106 may also include an echo canceller (not shown), which cancels sound echoes caused by feedback between a speaker (not shown) and earphone 102 .

After processing by speech processor 106, the signal is provided to transmitter 110, which performs modulation according to a preset format such as Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), or Frequency Division Multiple Access (FDMA). In the exemplary embodiment, transmitter 110 modulates signals according to the CDMA modulation format described in U.S. Patent No. 4,901,307, entitled "Spread Spectrum Multiple Access Communication System Utilizing Satellite or Ground-Based Repeaters," which is incorporated by reference it's here. The transmitter then upconverts and amplifies the modulated signal, and sends the modulated signal through antenna 112 .

It should be appreciated that noise suppressor 108 may be implemented within a speech processing system other than system 100 of FIG. 1 . For example, noise suppressor 108 may be used in an email application that includes a voicemail option. For such applications, the transmitter 110 and antenna 112 of FIG. 1 are no longer required. Instead the noise suppressed signal is formatted by speech processor 106 for transmission over the email network.

FIG. 2 shows an embodiment of the noise suppressor 108 . As shown in FIG. 2 , an input audio signal is received by a pre-processor 202 . The pre-processor 202 prepares the input signal for noise suppression by pre-emphasis and frame generation. Pre-emphasis redistributes the power spectral density of the speech signal by enhancing the high-frequency speech components of the signal. The pre-emphasis basically completes the high-pass filtering function, which strengthens the important speech components to improve the SNR of the three components in the frequency domain. Preprocessor 202 may also generate frames from input signal samples. In the preferred embodiment, 10 microsecond frames of 80 samples/frame are generated. Frames can contain overlapping samples for greater processing precision. Frames are generated by windowing and zeroing the samples of the input signal. The preprocessed signal is provided to a transform unit 204 . In a preferred embodiment, the transform unit 204 generates a 128-point Fast Fourier Transform (FFT) for each frame of the input signal. However, it should be understood that other means may be used to analyze the frequency components of the input signal.

The transformed components are provided to a channel energy estimator 206a, which produces energy estimates for each of the N transformed signal channels. For each channel, one technique for updating the channel energy is to estimate the smooth update of the current frame energy to the current frame channel energy as follows:

E _u (t)=αE _ch +(1-α)E _u (t-1) (1)

Here the updated estimate E _u (t) is defined as a function of the current channel energy E _ch and the previously estimated channel noise energy E _u (t-1). The embodiment sets α = 0.55.

The preferred embodiment determines an energy estimate for the low frequency channel and an energy estimate for the high frequency channel such that N=2. The low-frequency channel corresponds to a frequency of 250-2250 Hz, while the high-frequency channel is noisy at a frequency of 2250-3500 Hz. The current channel energy of the low-frequency channel can be determined by summing the FFT point energy corresponding to 250-2250 Hz, and the current channel energy of the high-frequency channel can be determined by summing the FFT point energy corresponding to 2250-3500 Hz.

The energy estimate is provided to speech detector 208, which determines whether speech is present in the received speech signal. SNR estimator 210a of speech detector 208 receives the energy estimate. SNR estimator 210a determines a speech signal-to-noise ratio (SNR) for each of the N channels based on the channel energy estimate and the channel noise energy estimate. The channel noise energy estimate is provided by the noise energy estimator 214a and typically corresponds to the estimated noise energy smoothed over previous frames that did not contain speech.

The voice detector 208 also includes a rate determination unit 212 that selects the data rate of the input signal from a set of preset data rates. In some communication systems, data is encoded such that the data rate can vary from frame to frame. This is called a variable rate communication system. A speech encoder that encodes data according to a variable rate scheme is generally referred to as a variable rate vocoder. An example of a variable rate vocoder is found in US Patent No. 5,414,796, entitled "Variable Rate Vocoder," which is incorporated herein by reference. Utilizing a variable rate communication channel eliminates unnecessary transmissions when there is no useful speech to send. Inside the vocoder, an algorithm is used to generate the rate at which the number of bits of information changes per frame, based on changes in speech activity. For example a vocoder with a set of four rates can generate 20 millisecond data frames containing 16, 40, 80 or 171 information bits depending on the activity of the speaker. Each data frame needs to be sent within a fixed time by varying the communication transmission rate.

Since the frame rate is dependent on the speech activity during the time frame, the determination of the rate provides information on whether speech is present or not. In systems utilizing variable rates, a determination of whether a frame should be encoded at the highest rate generally indicates the presence of speech, while a determination of whether a frame should be encoded at the lowest rate generally indicates the absence of speech. Moderate rates generally indicate transitions between the presence and absence of speech.

Rate decision unit 212 may be implemented with any number of rate decision algorithms. Such a rate decision algorithm is disclosed in U.S. Patent No. 5,911,128, entitled "Method and Apparatus for Reduced Variable Rate Vocoding," issued June 8, 1999, which is incorporated herein by reference . This technique provides a set of rate decision criteria called mode metrics. The first pattern metric is the target matching signal-to-noise ratio (TMSNR) from previously encoded frames, which provides information on how well the encoding model can be completed by comparing the synthesized speech signal with the input speech signal. The second pattern metric is the normalized autocorrelation function (NACF), which measures periodicity in speech frames. The third mode metric is the zero-crossing (ZC) parameter, which measures the high-frequency content within the input speech frame. The fourth mode metric is the Prediction Gain Difference (PGD), which determines whether the encoder maintains its predictive efficiency. The fifth mode metric is Energy Difference (ED), which compares the energy within the current frame to the average frame energy. Using these mode metrics, the rate decision logic selects an encoding rate for the incoming frames.

It should be understood that although FIG. 2 shows the rate determination unit 212 included as a unit of the noise suppressor 108, the rate information may also be provided to the noise suppressor 108 (FIG. 1) by another unit of the speech processor 106. For example, speech processor 106 may include a variable rate vocoder (not shown), which determines the encoding rate per frame of the input signal. Instead of noise suppressor 108 making the rate determination independently, the rate information may be provided to noise suppressor 108 by a variable rate vocoder.

It should be appreciated that instead of utilizing the rate judgment to determine the presence of speech, speech detector 208 may employ a subset of the mode metrics associated with the rate judgment. For example the rate decision unit 212 can be replaced by a NACF unit (not shown), which measures the periodicity within speech frames as described above. NACF is valued according to the following relationship:

$\mathrm{<span\; class="notranslate">\; NACF</span>}<span\; class="notranslate">\; =</span>\frac{\stackrel{\mathrm{<span\; class="notranslate">\; max</span>}}{\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; \{</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span>}{\mathrm{<span\; class="notranslate">\; 0.5</span>}<span\; class="notranslate">\; \&CenterDot;</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; \{</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Here N is the number of samples of the speech frame, and t1 and t2 are boundaries within T samples used to estimate NACF. NACF estimates NACF from the formant residual signal e(n). The formant frequency is the resonant frequency of speech. The speech signal is filtered with a short period filter to obtain the formant frequencies. The residual signal filtered by the short-period filter is a formant residual signal and contains long-period speech information, such as signal pitch.

Since the periodicity of a signal contained within a speech signal is different from that of a signal not contained within a speech signal, the NACF pattern metric is suitable for determining the presence or absence of speech. Speech signals are always characterized by periodic components. When speech is absent, the signal generally has no periodic component. The NACF metric is therefore a better indicator and can be used by the speech detector 208 .

Speech detector 208 may employ a metric such as NACF in place of a rate determination where no rate determination can be made. For example, if a rate call is not available from a variable rate vocoder, and the noise processor 108 does not have the processing capability to generate its own rate call, then a mode metric such as NACF provides the desired option. This could be an automotive application where processing power is constrained.

Furthermore, it should be understood that speech detector 208 may make a determination of whether speech is present based solely on rate determinations, mode metrics, or SNR estimates. While adding metrics should improve the precision of the judgment, a single metric already yields adequate results.

The rate decision (or mode metric) and the SNR estimate generated by the SNR estimator 210 a are provided to the speech decision unit 216 . The speech judgment unit 216 generates a judgment of whether speech is present in the input signal based on its input. The presence or absence of speech determines whether the noise energy estimate should be updated. The noise energy estimate is used by SNR estimator 210a to determine the SNR of speech in the input signal. SNR is used in turn to calculate the level of input signal attenuation for noise suppression. If it is determined that there is speech, the speech determination unit 216 turns on the switch 218a, preventing the noise energy estimator 214a from updating the noise energy estimate. If it is determined that there is no speech, the input signal is assumed to be noise, and the speech determination unit 216 closes the switch 218a, causing the noise energy estimator 214a to update the noise estimate. Although the switch 218a is shown in FIG. 2, it should be understood that the enable signal provided by the speech determination unit 216 to the noise energy estimator 214a can perform the same function.

In a preferred embodiment, what is estimated is the SNR of two channels, and the speech judgment unit 216 generates a noise update judgment according to the following procedure:

           
if(rate==min)

        if((chsnr1>T1)OR(chsnr2>T2))

            if(ratecount＞T3)

                 update noise estimate

            else

                 ratecount++

         else

            update noise estimate

            ratecount=0

    else

        ratecount=0

        

The channel SNR estimates provided by SNR estimator 210a are denoted chsnr1 and chsnr2. The rate of the input signal provided by the rate judging unit 212 is represented by rate. The counter, the rate count, tracks the number of frames based on certain conditions as described below.

The voice judging unit 216 judges that the voice does not exist and judges that the noise estimate should be updated. If the rate is the minimum rate in the variable rate, then chsnr1 is greater than the threshold T1 or chsnr2 is greater than the threshold T2, and the rate count is greater than the threshold T3. If the rate is minimum, and chsnr1 is greater than threshold T1 or chsnr2 is greater than threshold T2, but the rate count is less than threshold T3, then the rate count is incremented but the noise estimate is not updated. Counter, ie rate counting By counting frames with a minimum rate but high energy in at least one channel, a sudden increasing level of noise or a gradually increasing noise source is detected. A counter providing an indicator that a high SNR signal does not contain speech is set to count until speech is detected within the signal. A preferred embodiment sets T1 = T2 = 5dB, and T2 = 100 frames, where a frame estimate of 10 milliseconds is used.

If the rate is minimal, chsnr1 is less than T1, and chsnr2 is less than T2, then speech determination unit 216 will determine that speech is absent and the noise estimate should be updated. Also, the rate count is reset to zero.

If the rate is not minimum, the speech decision unit 216 will determine that the frame contains speech and not update the noise estimate, but reset the rate count to zero.

Instead of using a rate metric to determine the presence of speech, a pattern metric such as NACF can be employed. Speech judgment unit 216 may utilize NACF metrics to determine the presence of speech and noise update judgments according to the following procedure:

           
if(pitchPresent==FALSE)

        if((chsnr1>TH1)OR(chsnr2>TH2))

             if(pitchCount＞TH3)

                 update noise estimate

             else

                  pitchCount++

         else

              update noise estimate

              pitchCount=0

    else

         pitchCount=0

        

Here pitchPresent is defined as follows:

           
if(NACF>TT1)

        pitchPresent=TRUE

        NACFcount=0

    elseif(TT2≤NACF≤TT1)

        if(NACFcount＞TT3)

            pitchPresent=TRUE

        else

            pitchPresent=FALSE

            NACFcount++

    else

         pitchPresent=FALSE

         NACFcount=0

        

The channel SNR estimates provided by SNR estimator 210a are also denoted chsnr1 and chsnr2. A NACF unit (not shown) generates the metric pitchPresent as defined above indicating the presence or absence of a pitch. The counter, pitchCount, tracks the number of frames based on certain conditions as described below.

The metric pitchPresent determines that a pitch is present if NACF is greater than a threshold TT1. A tone is also determined to be present if NACF is in the middle range of several frame numbers greater than threshold TT3 (TT2≦NACF≦TT1). A counter, NACFcount tracks the number of frames for which TT2≤NACF≤TT1. In a preferred embodiment, TT1 = 0.6, TT2 = 0.4, and TT3 = 8 frames, where estimates are for 10 millisecond frames.

Speech judgment unit 216 judges that speech does not exist and should update the noise estimate, if the pitchPresent metric indicates that pitch does not exist (pitchPresent=False), then chsnr1 is greater than threshold TH1 or chsnr2 is greater than threshold TH2, and pitchCount is greater than threshold TH3. If pitchPresent = False, and chsnr1 is greater than TH1 or chsnr2 is greater than TH2, but pitchPresent is less than TH3, pitchPresent is incremented but the noise estimate is not updated. A counter, pitchCount, is used to detect a sudden increase in the level of noise or a gradually increasing noise source. The preferred embodiment sets T1 = T2 = 5dB, and T2 = 100 frames, where the estimate is a 10 ms frame.

If pitchPresent indicates that no pitch is present, and chsnr1 is less than TH1 and chsnr2 is less than TH2, then speech determination unit 216 will determine that speech is not present and the noise estimate should be updated. Also, pitchCount is reset to zero.

If pitchPresent indicates that pitch rate is present (pitchPresent=TRUE), speech decision unit 216 will determine that the frame contains speech and not update the noise estimate, but reset pitchCount to zero.

On the basis of judging that there is no speech, the switch 218a is closed to allow the noise energy estimator 214a to update the noise estimate. Noise energy estimator 214a typically generates noise energy estimates for each of the N channels of the input signal. Since speech is absent, it is assumed that the energy is all contributed by noise. For each channel, the noise energy update is estimated as the smoothing of the current channel energy over the channel energy of previous frames that did not contain speech. An updated estimate can be obtained, for example, from the following relationship:

E _u (t)=βE _ch +(1-β)E _u (t-1) (3)

Here the updated estimate E _u (t) is defined as a function of the current channel energy E _ch and the previously estimated channel noise energy E _u (t-1). The embodiment sets β=0.1. The updated channel noise energy estimate is provided to SNR estimator 210a. These channel noise energy estimates will be used to obtain an update of the channel SNR estimate for the next frame of the input signal.

A determination as to whether speech is present is also provided to channel gain estimator 220 . Channel gain estimator 220 determines the gain and noise suppression level for the input signal frame. If the voice determination unit 216 has determined that the voice does not exist, the frame gain is set to a preset minimum gain level. Otherwise, the gain is determined as a function of frequency. In the preferred embodiment, the gain is calculated from the graph of FIG. 3 . Although FIG. 3 is in graphical form, it should be understood that the function shown in FIG. 3 may be implemented in the form of a look-up table within channel gain estimator 220 .

It can be seen from FIG. 3 that the embodiment of the present invention defines a separate gain curve for each of the L frequency bands. Although L can be any number greater than or equal to 1, in FIG. 3 there are 3 frequency bands (L=3). Thus the gain factor of the low-band channel can be determined using the low-band curve, the gain factor of the mid-band channel can be determined using the mid-band curve, and the gain factor of the high-band channel can be determined using the high-band curve.

Although noise suppression can be accomplished using only one gain curve (L=1) of the input signal, speech quality degradation can be reduced by using multiple frequency bands. In ambient noise (such as road and windy conditions), the energy of the noise signal is higher at low frequencies and generally decreases with increasing frequency.

In Figure 3, a linear equation with fixed slope and y-intercept is used to determine the gain factor for each frequency band. The determination of the gain factor can be described by the following equation:

gain[low band](dB)=slope1*SNR+lowBandYintercept; (4)

gain[mid band](dB)=slope2*SNR+midBandYintercept; (5)

gain[high band](dB)＝slope3*SNR+highBandYintercept. (6)

The preferred embodiment specifies low frequencies from 125-375 Hz, mid frequencies from 375-2625 Hz, and high frequencies from 2625-4000 Hz. Slope and intercept were determined experimentally. Although a different slope can be used for each frequency band, the preferred embodiment uses the same slope of 0.39 for each frequency band. And lowBandYintercept is set to -17dB, midBandYintercept is set to -13dB, and highBandYintercept is set to -13dB.

The options feature will provide the user with the means to include a noise suppressor to select the desired y-intercept. Thus greater noise suppression (lower y-intercept) can be chosen at the expense of speech quality degradation. The y-intercept may be a variable of a function of some metric determined by the noise suppressor 108 . Stronger noise suppression (lower y-intercept) may be required, for example, when excess noise energy is detected within a predetermined time interval. Weaker noise suppression (higher y-intercept) may be required when conditions such as crosstalk are detected. During babble, background talkers are present and lower noise rejection may be warranted to prevent cutting off of the main talker. Another optional feature would provide a selectable gain curve slope. And it should be understood that, in addition to the curves described in equations (4)-(6), there may be other curves that are more suitable for determining the gain factor in a certain situation.

For each frame containing speech, a gain factor is determined for each of the M frequency channels of the input signal, where M is the predetermined number of channels being evaluated. The preferred embodiment evaluates 16 channels (M=16). Referring to FIG. 3, the low frequency curve is used to determine the gain factor of a channel having frequency components in the low frequency range. Use the IF curve to determine the gain factor for a channel with frequency components in the IF range. The high frequency curve is used to determine the gain factor for a channel with frequency components in the high frequency range.

For each estimated channel, the gain factor is derived from the appropriate curve using the channel SNR. The channel SNR shown in FIG. 2 is estimated by channel energy estimator 206b, noise energy estimator 214b, and SNR estimator 210b. For each frame of the input signal, the channel energy estimator 206b generates energy estimates for each of the M channels of the transformed input signal. The channel energy estimate can be updated using the relationship of equation (1) above. If the speech determination unit 216 determines that there is no speech in the input signal, the switch 218b is closed, and the noise estimator 214b updates the estimate of the channel noise energy. For each of the M channels, the updated noise energy estimate is based on the channel energy estimate determined by channel energy estimator 206b. The updated estimate can utilize the relational estimate of Equation (3) above. The channel noise estimate is provided to SNR estimator 210b. SNR estimator 210b thus determines a channel SNR estimate for each speech frame based on the channel energy estimate for the particular speech frame and the channel noise energy estimate provided by noise energy estimator 214b.

Those skilled in the art will recognize that channel energy estimator 206a, noise energy estimator 214a, switch 218a, and SNR estimator 210a perform functions similar to channel energy estimator 206b, noise energy estimator 214b, switch 218b and SNR estimator 210b complete the functions. Thus, although shown as separate processing units in FIG. 2, channel energy estimators 206a and 206b may be combined as one processing unit, noise energy estimators 214a and 214b may be combined as one processing unit, switches 218a and 218b may be combined as one unit, and the SNR estimators 210a and 210b can be combined as one unit. As a combining unit, the channel energy estimator will determine channel energy estimates for N channels for speech detection and M channels for determining channel gain factors. It is worth noting that the possible case is N=M. Likewise, the noise energy estimator and the SNR estimator will work on N channels and M channels. The SNR estimator then provides N SNR estimates to speech decision unit 216 and provides M SNR estimates to channel gain estimator 220 .

The channel gain factor is provided by channel gain estimator 220 to gain adjuster 224 . Gain adjuster 224 also receives the FFT transformed input signal from transform unit 204 . The gain of the converted signal is appropriately adjusted according to the channel gain factor. For example, in the above embodiment (where M=16), the transform (FFT) point belonging to one of the 16 channels is adjusted according to an appropriate channel gain factor.

The gain adjusted signal generated by gain adjuster 224 is then provided to an inverse transform unit 226 which, in the preferred embodiment, generates an inverse fast Fourier transform (IFFT) of the signal. The inverse transformed signal is provided to a post-processing unit 228 . If the input frame has been formed with overlapping samples, the post-processor unit 228 adjusts the overlapped output signal. If the signal has undergone pre-emphasis, the post-processing unit 228 also performs de-emphasis. De-emphasis attenuates the frequency separation that was emphasized during pre-emphasis. The pre-emphasis/de-emphasis process effectively suppresses noise by reducing noise components outside the frequency components to be processed.

It should be understood that the various processing blocks of the noise suppressor shown in FIG. 2 may be implemented as a digital signal processor (DSP) or an application specific integrated circuit (ASIC). A functional description of the invention will enable one of ordinary skill to implement the invention in a DSP or ASIC without undue experimentation.

See FIG. 4 for a flow diagram illustrating some of the steps involved in the processing described in FIGS. 2 and 3 . Although the steps are shown sequentially, those skilled in the art will recognize that the order of certain steps may be interchanged.

The process begins at step 402 . In step 404, the transform unit 204 transforms the input audio signal into a transformed signal, usually an FFT signal. In step 406, SNR estimator 210b determines speech SNRs for the M channels of the input signal based on the channel energy estimates provided by channel energy estimator 206b and the channel noise energy estimates provided by noise energy estimator 214b. In step 408, the channel gain estimator 220 determines the gain factors of the M channels of the input signal according to the channel frequencies. If there is no speech within the input signal frame, the channel gain estimator 220 sets the gain at a minimum level. Otherwise the gain factor for each of the M channels is determined according to a predetermined function. For example, referring to FIG. 3, a function defined by linear equations with fixed slope and y-intercept may be employed, where each linear equation defines a gain for a predetermined frequency band. In step 410, the gain adjuster 224 adjusts the gains of the M channels of the transformed signal using the M gain factors. In step 412, the inverse transform unit 226 inverse transforms the gain-adjusted transformed signal to produce a noise-suppressed audio signal.

In step 414, SNR estimator 210 determines speech SNRs for the N channels of the input signal based on the channel energy estimates provided by channel energy estimator 206a and the channel noise energy estimates provided by noise energy estimator 214a. In step 416, the rate determination unit 212 determines the encoding rate of the input signal by analyzing the input signal. Additionally, one or more pattern metrics such as NACF can be determined. In step 418 , speech determination unit 216 determines whether speech is present in the input signal based on the SNR provided by SNR estimator 210 , the rate and/or mode metric provided by rate determination unit 212 . If at decision block 420 it is determined that speech is not present, then the input signal is assumed to be completely noise and a noise estimate update is done at step 422 by the noise energy estimator 214a. Noise energy estimator 214a updates the noise estimate based on the channel energy determined by channel energy estimator 206a. No matter whether the voice is detected or not, the program continues to process the next signal frame.

The invention has been described above by means of the embodiments. Various modifications to the present invention will occur to those skilled in the art without creative effort. It is therefore intended that the scope and spirit of the invention be defined by the appended claims.

Claims (32)

1. A noise suppressor for suppressing audio signal background noise, characterized in that it comprises: a signal-to-noise ratio (SNR) estimator for generating channel SNR estimates for a first predefined set of frequency channels of the audio signal; a gain estimator for generating a gain factor for each of said frequency channels based on a corresponding one of said channel SNR estimators, wherein said gain factor is derived using a gain function defining the gain factor as an SNR increasing function; a gain adjuster for adjusting the gain level of each of said frequency channels according to one of said corresponding gain factors; and A speech detector for determining the presence of speech in the audio signal, wherein the speech detector utilizes an estimator and a rate decision unit to detect the presence of speech. 2. The noise suppressor of claim 1, wherein said gain function is frequency dependent. 3. The noise suppressor according to claim 1, characterized in that said gain function is implemented as a look-up table. 4. The noise suppressor of claim 1, wherein said gain function is a linear function with a fixed slope and y-intercept. 5. The noise suppressor of claim 4, wherein said y-intercept is user selectable. 6. The noise suppressor of claim 4, wherein the y-intercept is adjusted based on a measured characteristic of noise within the audio signal. 7. The noise suppressor of claim 4, wherein said slope is user selectable. 8. The noise suppressor of claim 4, wherein said slope is adjusted based on a measured characteristic of noise within said audio signal. 9. The noise suppressor of claim 1, further comprising: a noise energy estimator for generating an updated channel noise energy estimate for each of said frequency channels when said speech detector determines that there is no speech within said audio signal, said updated channel noise energy estimate being provided to said an SNR estimator to generate the channel SNR estimate. 10. The noise suppressor of claim 9, wherein said speech detector comprises: a signal-to-noise ratio (SNR) estimator for generating channel SNR estimates for a second predefined set of frequency channels of the audio signal; A voice judging unit, configured to determine whether there is voice according to the estimated channel SNR of the second frequency channel group. 11. The noise suppressor as claimed in claim 10, wherein said voice detector further comprises: a mode metric unit for determining at least one mode metric characterizing said audio signal; Wherein the speech determination unit determines the presence of speech based on the at least one pattern metric. 12. The noise suppressor of claim 11, wherein the pattern metric comprises a normalized autocorrelation function (NACF) metric. 13. A noise suppressor for suppressing background noise of an audio signal, characterized in that it comprises: means for detecting an encoding rate associated with said audio signal, wherein said audio signal is encoded according to an encoding rate; means for determining whether speech is present in said audio signal based on a coding rate; means for generating a channel signal-to-noise ratio (SNR) estimate for a predefined set of frequency channels of the audio signal; means for determining a gain factor for each of said frequency channels if said means for determining whether speech is present in said audio signal determines that speech is present, wherein a gain function is defined for each of a set of frequency bands, and for each of said The frequency band defines a gain factor that increases as the SNR increases, and the channel gain factor is determined according to a gain function that ranges over a frequency band that includes the frequency channel; and means for adjusting the gain level of each of said frequency channels according to said corresponding channel gain factor. 14. A noise suppressor as claimed in claim 13, wherein said means for determining a gain factor determines a minimum gain factor for each of said frequency channels if said means for determining whether speech exists determines that speech does not exist . 15. The noise suppressor of claim 13, wherein the gain function is implemented as a look-up table. 16. The noise suppressor of claim 13, wherein said gain function is a linear function with a fixed slope and y-intercept. 17. The noise suppressor of claim 16, wherein each of said y-intercepts is user selectable. 18. The noise suppressor of claim 16, wherein each of said y-intercepts is adjusted based on a measured characteristic of noise within said audio signal. 19. The noise suppressor of claim 16, wherein each of said slopes is user selectable. 20. The noise suppressor of claim 16, wherein each of said slopes is adjusted based on a measured characteristic of noise within said audio signal. 21. The noise suppressor of claim 13, further comprising: for generating an updated channel noise energy estimate for each of said frequency channels when said means for determining whether speech is present determines that there is no speech in said audio signal, said updated channel noise energy estimate being provided to a device for generating an SNR estimate Value means to update the channel SNR estimate. 22. The noise suppressor of claim 13, wherein said means for determining whether speech is present further comprises: means for generating channel SNR estimates for a second predefined set of frequency channels of the audio signal. 23. The noise suppressor of claim 13, wherein said means for determining whether speech is present comprises: means for determining at least one pattern metric characterizing said audio signal; and Means for determining the presence or absence of speech based on said at least one pattern metric. 24. The noise suppressor of claim 23, wherein said means for determining whether speech is present further comprises: means for generating channel SNR estimates for a second set of predefined frequency channels of the audio signal; The device for making a judgment on whether the voice exists further makes a judgment based on the SNR estimate. 25. The noise suppressor of claim 23, wherein said pattern metric comprises a normalized autocorrelation function (NACF) metric. 26. A method for suppressing background noise in an audio signal, comprising the steps of: transforming the speech signal into a frequency representation of the audio signal; detecting a coding rate associated with said audio signal; determining whether speech is present in the audio signal according to the encoding rate of the audio signal; generating channel signal-to-noise ratio (SNR) estimates for a predefined set of frequency channels represented by said frequency; If it is determined that speech is present within the audio signal, then determining a gain factor for each of said frequency channels, wherein a gain function is defined for each of a set of frequency bands, and for each of said frequency bands a gain factor that increases with increasing SNR is defined a gain factor, whereby for each of said frequency channels, the channel gain factor is determined according to a gain function ranging over a frequency band comprising the frequency channel; adjusting the gain level of each of said frequency channels according to said corresponding channel gain factor; and The gain-adjusted frequency representation is inverse transformed to produce a noise-suppressed audio signal. 27. The method of claim 26, comprising the steps of: A minimum gain factor for each of said frequency channels is determined if speech is determined to be absent. 28. The method of claim 26, wherein each of said gain functions is a linear function with a fixed slope and y-intercept. 29. The method of claim 26, further comprising the steps of: generating an updated channel noise energy estimate for each of said frequency channels when said step of determining whether speech is present determines that speech is absent within said audio signal, said updated channel noise energy estimate being used to generate said channel SNR estimate . 30. The method of claim 26, wherein the step of determining whether the voice exists comprises: generating channel SNR estimates for a second predefined set of frequency channels of the audio signal; Determining whether speech exists according to the channel SNR estimate of the second group of frequency channels. 31. The method of claim 30, wherein the step of determining whether the voice exists further comprises: determining at least one mode metric characterizing the audio signal; and A determination of whether speech is present is determined based on the at least one pattern metric. 32. The method of claim 31, wherein the pattern metric comprises a normalized autocorrelation function (NACF) metric.

Images